Ranking
Safest AI Chatbots and Assistants 2026
Every AI chatbot we cover, from general assistants such as ChatGPT, Claude, Gemini, Copilot and DeepSeek to companion and role-play apps such as Character.AI, Replika and zeta. Each score weighs what the app keeps from your chats and who sees it, how well it protects your account, the controls it gives you, how it handles minors and harmful answers, and its record with regulators.
Updated · Scores update with each page
The ranking
AI assistants and chatbots
Claude
Claude is reasonable for everyday questions, but it is not a private notebook. Anthropic trains on consumer chats unless you opt out, keeps them up to five years when training is on, and signs you in only through Google or an emailed link. Shared chats turned up on Google in July 2026.
Score 63, grade CAI assistants and chatbots
Google Gemini
Gemini sits on Google's strong account security and does not use your chats for ads. The trade-off is retention. Chats train Google's models unless you turn activity off, and conversations picked for human review are kept for up to three years, even if you delete them.
Score 61, grade CAI assistants and chatbots
Perplexity
Perplexity is a reasonable way to look things up, with an easy training opt-out and incognito mode. Searches train its models by default, the app can link your precise location to you, and its Comet browser, which can act inside your accounts, has been tricked by instructions hidden in web pages.
Score 59, grade CAI assistants and chatbots
Microsoft Copilot
Copilot is a reasonable choice for everyday questions. Since August 2026 Microsoft says the new app does not train its models on your chats, but the app tracks you for advertising, ads can be personalised with your conversations, and a one-click data-theft flaw took eight months to fix.
Score 58, grade CAI assistants and chatbots
ChatGPT
ChatGPT is secure enough for everyday questions, but it is not a private diary. OpenAI stores your chats, uses them for training by default, shows ads to free users in the US, and faces lawsuits claiming it failed vulnerable teenagers.
Score 57, grade CAI assistants and chatbots
Muse from Meta
Muse is Meta's adults-only AI agent that reads your email, books, buys and negotiates for you. Meta built strong guard rails around it and promises your chats stay out of its ad systems, but the app is seven weeks old, has already had three security problems, and holds more of your life than any chatbot.
Score 53, grade DAI assistants and chatbots
Dola
Dola, formerly Cici, is ByteDance's AI assistant for markets outside China. It has no known breach or fine and lets you switch off model training, but it tracks you across apps, infers traits such as age and gender, and shares your data with ByteDance group companies and outside AI providers such as Google.
Score 49, grade DAI assistants and chatbots
Character.AI
Character.AI is an adult role-play app with a troubled record on teenagers. It tracks users for advertising, keeps every chat, and faces lawsuits from Kentucky and Pennsylvania's medical board. Since November 2025 under-18s can no longer chat freely with its characters.
Score 46, grade DAI assistants and chatbots
Doubao
Doubao is China's most used AI assistant, made by TikTok's owner ByteDance. It has no known breach or fine, and it lets you turn off training and delete chats, but your data stays in China under Chinese law, and the app tracks you with identifiers and collects precise location.
Score 46, grade DAI assistants and chatbots
Grok
Grok's app collects less than most AI apps and lets you opt out of training, but its safety record is the weakest of the big chatbots. In early 2026 it was used to make sexualised images of real people, including children, and regulators in the EU, Ireland and the UK opened investigations that are still running.
Score 45, grade DAI assistants and chatbots
zeta
zeta is Scatter Lab's AI role-play app, popular with teenagers in Korea and Japan. Its maker was fined in 2021 for training a chatbot on private KakaoTalk chats, zeta trains on your chats by default, and a 2026 test found its teen-safe mode could still produce sexual abuse scenarios.
Score 44, grade DAI assistants and chatbots
Meta AI
Meta AI is built into Facebook, Instagram, WhatsApp and its own app, and it feeds Meta's ad business. Outside the EU, UK and South Korea, what you tell it shapes your ads, with no switch to stop it. Its record with teenagers and a 2025 privacy bug add to the concerns.
Score 43, grade DAI assistants and chatbots
Replika
Replika is an adults-only AI companion that stores some of the most personal conversations any app holds, yet we found no two-factor login. Italy's privacy regulator fined its maker €5 million in 2025 and found its age checks still deficient, and a US complaint alleges it is designed to build emotional dependence.
Score 43, grade DAI assistants and chatbots
DeepSeek
DeepSeek's chatbot is free and capable, but everything you type is stored on servers in China. Regulators in Italy, South Korea and Berlin have acted against its data transfers, and in January 2025 researchers found a database of user chats open on the internet. Use it only for questions you would happily publish.
Score 32, grade F
Which apps are on this list
The rule is simple: the app's main feature is a conversation with an AI model. That covers the general assistants (ChatGPT, Claude, Google Gemini, Microsoft Copilot, Perplexity, Grok, Meta AI and DeepSeek), ByteDance's Doubao for China and Dola for other markets, and Meta's Muse, an agent that books and buys for you on top of the chat. Companion and role-play apps (Character.AI, Replika and zeta) are on the list too. They are chatbots as well, and the questions people ask about AI, such as what happens to intimate chats and what a bot says to a teenager, are sharpest there.
The music generators Suno and Udio sit in the same AI section of our app list but are left out, because you make songs with them rather than talk to them. Chatbots built into other apps, such as My AI in Snapchat, are scored on that app's page.
How the order is built
Each app is scored on five factors. Privacy and security count for 30% each. For a chatbot, privacy mostly means what happens to your conversations: how long they are kept, whether they train models and whether they feed ads. Transparency and control count for 15%, scams and harmful content for 15% (age checks, crisis responses and what the model produces when pushed), and the regulatory record for 10%. The full rubrics are on how we score apps.
The list is those scores sorted from the highest down. When we update a page, its place here moves too, so this text does not quote scores.
What the apps near the top have in common
No chatbot does well on privacy, so the top of this list is a set of reasonable compromises rather than private tools. The general assistants from Anthropic, Google, Perplexity, Microsoft and OpenAI sit within a few points of each other there. Claude, Gemini and Perplexity have App Store labels with no tracking, offer a clear training opt-out, and have no feed of strangers' posts. Gemini also sits behind Google's account security, with PasskeyA way to sign in without a password. Your device holds a key tied to one website or app and unlocks it with your fingerprint, face or PIN.More in the glossary and the Advanced Protection Program, while Claude still signs you in only through Google or an emailed link. None of the three has a privacy fine of its own. Copilot no longer trains on chats in its new app but tracks you for advertising, and ChatGPT shows ads to free users in the US.
Why some well-known chatbots score lower
Meta AI has used what you tell it to shape ads since December 2025 outside the EU, UK and South Korea. Grok was used in early 2026 to make sexualised images of real women and children, and the European Commission, Ireland's privacy regulator and Ofcom opened investigations. DeepSeek stores everything in China, left a database of chats open in 2025, and in the same year Italy's privacy regulator blocked it and South Korea's suspended new downloads for about two months.
The companion apps lose most on harmful content. Their users share their most personal thoughts, and their records with minors are weak. Italy fined Replika's maker €5 million in 2025, and we found no Two-factor authenticationSigning in with a second proof of identity on top of your password, such as a code from an app, a security key or a passkey.More in the glossary login on the app. Character.AI faces a lawsuit from Kentucky's attorney general and filed for bankruptcy in October 2026, leaving open what happens to stored chats if the company is sold.
What the score does not tell you
The score does not measure whether answers are right. Every chatbot here can state wrong health, legal or money advice with confidence, so check anything important with a person or an official source. Do not type passwords, card numbers, ID numbers or confidential work documents into a consumer chatbot. Be careful with share links as well: Google indexed more than 370,000 shared Grok chats in 2025, and shared Claude chats turned up in Google results in July 2026. Before you share a chat, check who can open the link.
Questions
Is ChatGPT safe to use?
It is secure enough for everyday questions, but it is not private. OpenAI stores your chats without End-to-end encryptionEncryption in which only the sender and the recipient can read a message. The company running the service cannot see its content.More in the glossary, trains its models on them by default on the Free, Go and Plus plans, and has shown ads to logged-in free users in the US since February 2026. Several families have sued OpenAI alleging the chatbot failed teenagers in crisis, which OpenAI denies. Turn off "Improve the model for everyone" in the data controls if you do not want your chats used for training.
Is DeepSeek safe?
Use it only for questions you would be happy to see published. DeepSeek's privacy policy says your prompts, files and chat history are stored in China, and in January 2025 the security firm Wiz found an open DeepSeek database holding user chat history. Italy ordered it to stop processing Italians' data that month, and Australia, Taiwan and South Korea ban it on government devices. Running DeepSeek's open model on your own computer, or through a provider outside China, keeps your prompts off its servers.
Which AI chatbot is the most private?
None is private in the way an encrypted messenger is. Every chatbot on this list keeps your chats on its servers without End-to-end encryptionEncryption in which only the sender and the recipient can read a message. The company running the service cannot see its content.More in the glossary, and staff can review them. The App Store labels for Claude, Gemini and Perplexity show no tracking, but all three train on chats unless you opt out. Copilot's new app stopped training on chats in August 2026, yet it tracks you for advertising. Meta AI uses your chats to shape ads outside the EU, UK and South Korea, with no switch to stop it.
Do AI chatbots train on my chats?
Most do by default. ChatGPT, Claude, Gemini, Perplexity, Grok, Muse, Dola and zeta all use your conversations to improve their models unless you switch it off, and zeta only lets you opt out by email. Microsoft says the new Copilot app does not train on chats. Opting out matters for how long data is kept too. Claude keeps chats for up to five years when training is on, and Gemini keeps chats picked for human review for up to three years, even after you delete them.
Are AI companion apps safe for teenagers?
We would not let a teenager use one unsupervised. Character.AI ended open-ended chat for under-18s in November 2025, after families sued over harm to teenagers, including a 14-year-old who died by suicide in 2024. Replika is for adults, but Italy's privacy regulator found in 2025 that its age checks were still deficient. About 380,000 of zeta's 1.48 million monthly users in Korea were 19 or younger in September 2026, and a lawmaker's test found its teen mode still produced sexual violence scenarios. Use the phone's age limits to block apps rated above your child's age.